Compute-region admission alternatives — 2026-08-10

Verdict

INCOMPLETE. This packet executes fak’s region admission and an unbounded/no-admission baseline. Kubernetes, Slurm, Ray, and AWS Batch have zero measurements until real schedulers enforce the same region inventory; #6139 tracks those witnesses.

Same-workload contract

Every arm receives one exclusive live lease over device region 0-1, a declared device address space 0-7, and four claims: overlapping 1-2, disjoint 2-3, out-of-taxonomy 8, and different-class KV tier 0. The oracle requires collision refusal, disjoint admission, taxonomy refusal, and cross-class admission. Full witnesses report admission quality, constraint violations, scheduling latency/throughput, CPU/RSS/control-plane bytes, accelerator idle time, and total cost.

Arm	Class	Local availability	Honest result
fak native compute-region admission	native	yes	correct four-decision sequence
dispatch without region admission	tuned no-feature baseline	yes	incorrect: admits collision and taxonomy violation
Kubernetes scheduler	external	no	zero measurements
Slurm scheduler	external	no	zero measurements
Ray scheduler	external	no	zero measurements
AWS Batch	external	no	zero measurements

Repository inspection found no equivalent first-class fak integration distinct from this native admission seam. Configuration adapters and mocks are not scheduler witnesses.

Local native witness

go test ./internal/computeadmit -bench BenchmarkDecideComputeRegionAdmission -benchmem -run '^$' -count=5

Windows/amd64, AMD Ryzen 9 9950X: 1107, 1120, 1854, 1757, 1696 ns/op. Median: 1696 ns/four decisions (about 424 ns/decision), 1105 B/op, 36 allocs/op. This is pure policy execution, not end-to-end scheduler latency or accelerator cost.

Reproduce

go test ./internal/computeadmit -run TestCompareLocalKeepsSchedulerAlternativesExplicit
go test ./internal/computeadmit -bench BenchmarkDecideComputeRegionAdmission -benchmem -run '^$' -count=5
go test ./internal/nativebench